9 of 9 people found the following review helpful
Stephen E. Robbins
- Published on Amazon.com
Format: Kindle Edition
This book is an extensive treatment of the thesis that the brain employs Bayesian prediction to determine our perception of the world. Hohwy has obviously thought about his subject deeply, illustrates/explains the principles well, shows great familiarity with the group of thinkers (e.g., Eliasmith, Friston, Gregory) that are main sources of this view, and produces a remarkable effort at extending this principle through various areas to which it could apply. These areas include extremely interesting explications of various illusions, particularly a class of illusions involving, or variants of, the "rubber hand," also inattentional blindness, attention, as a possible causal factor in delusions and mental illness, and more. Within are some remarkable implications for the nature of the "self" and our experience thereof. It is a conception and a work I consider to have a deep core of essential merit and is an essential read.
In this framework, following its early originator, Helmholtz, "...perceptions are regarded as similar to the predictive hypotheses of science, but are psychologically projected into external space and accepted as our most immediate reality." Thus the "problem of perception," as Hohwy argues at the start of the book, is how the right hypothesis about the world is shaped and selected. The brain is seen as maintaining a generative model of the world, an internal mirror of nature that recapitulates the causal structure of the world, and prediction error is minimized relative to the model's expected (hypothesized) states. In an inversionary twist, perceptual inference is always trying to use its prior knowledge to predict and suppress (yes, suppress) the sensory input the system is receiving. Thus, rather than the "bottom-up" data (received from the external world) being used as the material to create the perception, it is the predictive model best fitting or minimizing the error that is actually perceived.
Here I must add some critical observations. What this internal predictive model could actually look like, embodied as it must be in chemical-neural flows, how it could possibly look (in the brain) like my experience of the kitchen with its table, its chairs, and my hand with spoon stirring coffee in a cup on the table, and how this internal model is "projected into space," i.e., how, from this internal, neural-based "hypothesis" we obtain an image of the external world - these are questions one instantly (and impatiently) presumes that the book will quickly deliver on. But in an unfortunate structural aspect of the book, we wait until about 200 pages in to hit the disclaimer that, "this is not intended as a proposal that can explain why perceptual states are phenomenally conscious rather than not," that this is merely, "a proposal that describes the states that are conscious... [via] those representations of the world that currently best predict the sensory input," and that it "does not intend to touch the hard problem" (i.e., Chalmers' hard problem of consciousness).
Clearly, this Bayesian framework is at best a partial answer; it is a computational piece of the puzzle, but just a computational piece. There is no true, concrete dynamics involved - where by concrete dynamics, I mean a dynamics as concrete as that of an AC motor generating a field of force. Computations alone cannot account for consciousness. While clearly aware of this, at least to some extent, Hohwy yet tries to take the framework into explaining Searle's Chinese Room, i.e., how such a Bayesian network could account for the conscious understanding of an event which is created via the mediating device of a string of linguistic symbols - "The man stirred the coffee with the spoon." But he does not acknowledge that beneath this sentence-created event perception, lies the same difficulty as that beneath the hard problem of explaining the origin of the image of the kitchen, chairs and coffee-stirring spoon.
Events. While Hohwy describes the earlier-mentioned "causal structure of the world" (captured/recapitulated in these hypotheses) in terms of invariance (regularities) which exist over various scales of time, the discussion of these is very abstract, the concept of invariance seeming very limited - almost entirely in terms of "causal regularities," e.g., dropping an egg to the floor => a broken egg. Missing is any reference to ecological psychology and J. J. Gibson, a discipline and theory where the regularities have a precise mathematical structure - texture gradients, gradients over velocity flows, tau ratios, adiabatic ratios, inertial tensors, etc. Strangely missing too is any reference to (or attempt at integration with) the premier Bayesian model of the perception of dynamic form (Weiss, Simoncelli and Adelson, Nature Neuroscience, 2002), a model simply and concretely employing mathematically specified constraints (priors) upon estimates of the optical velocity flows of Gibson (for a review, "On Time, Memory and Dynamic Form," Consciousness and Cognition, 2004). All of these - the flow fields, inertial tensors, adiabatic ratios, etc - comprise the structure of invariance defining even the little event of "coffee stirring." It is a structure where the invariants are defined over time - an extended, flowing time. As Gibson argued, these invariants cannot exist in an "instant" or be transmitted as "bits" over the nerves. It is this dynamic structure over time that is arriving in the brain as the "bottom up" information - to be compared against an operative hypothesis and suppressed. And...it is this very same dynamic structure that would have to be incorporated within - would have to define - this operative "hypothesis" of the stirring event that is being "projected" as our experience. What would the "matching" or comparison process (this is in reality a "comparator" model) of these two dynamic structures (hypothesis vs. the invariance structure of the dynamically changing external event) possibly look like? How are such structures - intrinsically dynamic flows - stored in the brain? Or can they be? How are such events retrieved, memory-wise, to become "an hypothesis?"
The invariance laws/structure defining events, the deep problem of the elementary memory that supports our perception of time flowing events such as the stirring spoon (which is a problem of the nature of time itself), the problem of storage of such events in the brain (there is no actual theory of event storage), the mechanism for their retrieval - these are just some of many subjects that need to be addressed in this conception. In a word, the predictive brain, despite its promise, or indeed because of it, will need to get very serious about the actual nature of its own hypotheses.